1. No category

EI @ Haas WP 253R
Does Better Information Lead to Better Choices?
Evidence from Energy-Efficiency Labels
Lucas Davis and Gilbert Metcalf
Revised March 2015
Energy Institute at Haas working papers are circulated for discussion and comment purposes.
They have not been peer-reviewed or been subject to review by any editorial board.
© 2014 by Lucas Davis and Gilbert Metcalf. All rights reserved. Short sections of text, not to
exceed two paragraphs, may be quoted without explicit permission provided that full credit is
given to the source.
http://ei.haas.berkeley.edu
Does Better Information Lead to Better Choices?
Evidence from Energy-Efficiency Labels
Lucas W. Davis*
UC Berkeley and NBER
Gilbert E. Metcalf
Tufts University and NBER
March 2015
Abstract
Information provision is a key element of government energy-efficiency
policy, but the information that is provided is often too coarse to allow
consumers to make efficient decisions. An important example is the
ubiquitous yellow “EnergyGuide” label, which is required by law to be
displayed on all major appliances sold in the United States. These labels
report energy cost information based on average national usage and
energy prices. We conduct an online stated-choice experiment to measure
the potential welfare benefits from labels tailored to each household’s state
of residence. We find that state-specific labels lead to significantly better
choices. Consumers choose to invest about the same amount overall in
energy-efficiency, but the allocation is much better with more investment
in high-usage high-price states and less investment in low-usage low-price
states. The implied aggregate cost savings are larger than the cost of
implementing state-specific labels.
Key Words: Energy-Efficiency, Inattention, Information Provision, Energy Demand, EnergyGuide
JEL: D12, H49, Q41, Q48
* (Davis) Haas School of Business, University of California, Berkeley; Energy Institute at Haas; and National Bureau of
Economic Research. Email: [email protected]. (Metcalf) Department of Economics, Tufts University and National
Bureau of Economic Research. Email: [email protected]. We are grateful to Hunt Allcott, Stefano DellaVigna, Sébastien
Houde, and numerous seminar participants for helpful comments. We also thank Poom Nukulkij at GfK for his work on the
experiment. Data collected by Time-sharing Experiments for the Social Sciences, NSF Grant 0818839, Jeremy Freese and James
Druckman, Principal Investigators. This RCT was registered in the American Economic Association Registry for randomized
control trials under Trial number 560. The authors have not received any financial compensation for this project nor do they
have any financial relationships that relate to this research.
1.
Introduction
Information provision is a key element of government energy-efficiency policy.
An important example is the ubiquitous yellow “EnergyGuide” label, which is required by
law to be displayed on all major appliances sold in the United States. Similarly, new cars
and trucks sold in the United States must display information about vehicle fuel efficiency
and an estimate of annual gasoline expenditures. Over 40 countries worldwide have some
sort of energy-efficiency labelling requirements, (CLASP, 2014) .
This information is intended to help consumers make better decisions. However,
in many cases government-mandated labels do not provide accurate information
necessary for consumers to make efficient decisions. In particular, most labels report only
very coarse information based on national average energy prices and typical national
usage. In practice, energy prices and typical usage vary substantially, so the labels provide
information that is highly inaccurate for many consumers.
The objective of our project is to evaluate the potential welfare benefits from
providing more accurate information. We focus on room air conditioners because they
are a particularly lucid example. Within the lower 48 U.S. states we show that annual
cooling hours range by a 9:1 ratio, while electricity prices vary by more than a 2:1 ratio.
As a result, operating costs vary widely, from $28 per year in Washington state, to $316
per year in Florida. Despite these enormous differences in operating costs, consumers in
all states see the exact same EnergyGuide label.
We designed and implemented an online stated-choice experiment to measure
how consumer decisions would change with information tailored to each household’s
state of residence. We find that better labels indeed lead to better choices. When
presented with more accurate information, the average energy-efficiency of selected air
conditioners stays about the same, but the allocation is much better. Households facing
1
low energy prices and low expected usage invest less in energy-efficiency, while
households facing high energy prices and high expected usage invest more.
This
reallocation leads to lower lifetime costs – defined as the sum of purchase price plus
present discounted value of energy costs – for both types of households.
The implied aggregate savings are substantial. State-specific labels decrease
lifetime cost by an average of $10 per air conditioner, with larger savings in high-cost
states. U.S. consumers purchase more than 4 million room air conditioners each year, so
the implied aggregate cost savings exceed $40 million annually. Moreover, our results
suggest that state-specific labels would improve decision-making not just for room air
conditioners, but for a whole host of residential appliances.
We then provide additional analysis and evidence aimed at better understanding
the mechanisms underlying our results. We find that immediately after the experiment
most participants are unable to correctly answer basic questions about the information
they have just seen. Most do not know whether the labels they just saw were based on
national or state energy prices, nor do they know how energy prices or appliance usage in
their state compares to the national average.
Overall, the evidence points to people taking the information in these labels as
given without analyzing it carefully. Daniel Kahneman (2011) has referred to this kind of
decision making as WYSIATI, “What you see is all there is.” The content of the labels
changes participants’ decisions, so it is not that they are ignoring this information
completely. But they appear not to be exerting the additional effort that would be
required to understand what this information means nor are they spontaneously
transforming this information to take local conditions into account.
Our paper differs from most previous studies of energy-efficiency. While there is
an extensive theoretical and empirical literature on the economic determinants of
2
investments in energy-efficient capital, there is little that has taken an explicit
experimental approach. None of the work to date has focused on the efficiency cost of
inaccurate information provided to consumers as this study does.1,2 More generally, our
paper complements a growing broader literature that shows that customized information
can significantly improve education, health, and finance-related choices (see, e.g.,
Hastings and Weinstein, 2008; Bertrand and Morse, 2011; Kling et al., 2012; and Hoxby
and Turner, 2013).
It is worth emphasizing that our evidence comes from a stated-choice experiment.
The highly-stylized setting allows us to eliminate many of the factors that complicate
these decisions in real-world settings. This facilitates analysis and interpretation, but it
also may lead participants to focus more on labels than they otherwise would. One
approach to validating our results is to look for complementary evidence from actual
choices. Examining data from appliance purchases, we find a negative correlation
between operating costs and investments in energy-efficiency. Although this doesn’t tell
us how much choices would change with better information, it corroborates other results
in the paper about the lack of effectiveness of current labels.
The paper proceeds as follows. Section 2 provides background information, and
makes the case for why better information might matter. Section 3 describes our online
experiment. Sections 4 and 5 provide the main results and additional analysis. Section 6
offers concluding comments.
1
Studies focusing on consumer choice of energy efficient capital include Hausman (1979), Dubin and
McFadden (1984), Metcalf (1994), Revelt and Train (1998), Metcalf and Hassett (1999), and Davis (2008)
among others. See Gillingham, Newell and Palmer (2009) and Gillingham and Palmer (2014) for recent
surveys.
2
Two related studies perform online experiments using the same nationally-representative panel that we
employ. Newell and Siikamaki (2014) analyze optimal EnergyGuide label design while Allcott and
Taubinsky (Forthcoming) measure the effect of information provision on willingness-to-pay for energyefficient lightbulbs. Neither study considers the role of inaccurate information provided to consumers.
3
2.
Background
A.
Previous Research
Economists have long been interested in how consumers make energy-related
decisions. Hausman (1979) and Dubin and McFadden (1984) model durable good
purchase decisions as a household production problem in which there is a tradeoff
between purchase price and operating cost. Following these seminal studies, much of the
literature has focused on whether or not consumers undervalue operating cost when
making these tradeoffs (see, e.g., Metcalf 1994 and Metcalf and Hassett 1999). The most
recent evidence comes from vehicle purchases and indicates that consumers do not
undervalue (Busse, Knittel, and Zettelmeyer, 2013) or modestly undervalue operating
costs (Allcott and Wozny, 2014).
Another more recent strand in the literature has aimed at understanding specific
behavioral biases in energy-related decisions. Allcott (2013) and Allcott (2013) examine
“MPG Illusion’’, the idea that consumers may not understand the non-linear relationship
between miles-per-gallon and motor vehicle fuel consumption. Camilleri and Larrick
(2014) test whether vehicle preferences are affected by the scale in which fuel economy
information is expressed, for example, gallons per 100 miles versus gallons per 1000
miles. Lastly, Allcott and Taubinsky (Forthcoming) and Allcott and Sweeney (2015) test
for biased beliefs and imperfect information by measuring the effect of information
provision on demand for energy-efficient lightbulbs and hot water heaters, respectively.
There are also studies that examine the effect of environmental messaging like
Energy Star Certification (e.g. Newell and Siikamaki (2014); Houde (2014b)) and
“normative” letter grades for the energy-efficiency characteristics of products (e.g.
Brounen and Kok (2011)).3 The evidence shows that people respond to these non-price
3
Newell and Siikamaki (2014) is similar to our study in that it uses an online stated-choice experiment to
4
interventions though it is not always clear if this is because they trigger “warm glow”
responses or because they are indirectly providing information about private costs.
Finally, there is a group of papers that study the effect of peer comparisons.
Learning about how your electricity consumption compares to your neighbors tends to
significantly reduce consumption, both in the short-run and long-run. See, for example,
Ayres, Raseman and Shih (2009), Allcott (2011b), and Allcott and Rogers (Forthcoming).
We see what we are doing as quite different. We are not studying consumers’
undervaluation of energy costs, nor are we studying a specific behavioral bias like MPG
illusion. Moreover, we have designed our experiment explicitly to exclude any
environmental messaging or peer comparisons. Instead, we are focused sharply on the
quality of the information that is publicly provided, and we want to ask whether better
tailoring this information to consumers’ characteristics can lead to more efficient choices.
B.
U.S. Energy Labeling Requirements
EnergyGuide labels must be displayed on all major appliances sold in the United
States. As of 2015, this includes clothes washers, dishwashers, refrigerators, freezers,
televisions, water heaters, window air conditioners, central air conditioners, furnaces,
boilers, heat pumps, and pool heaters. Collectively, these appliances account for over 60%
of residential energy consumption, and 13% of total U.S. energy consumption.4
Energy efficiency labels have existed since the first energy crisis in the mid-1970s.
evaluate components of EnergyGuide labels. In addition to comparing choices with and without Energy Star
certification, they randomly include or exclude information about carbon dioxide emissions, normative
letter grades, and other elements of label design. They do not, however, vary the operating cost information
itself or explore information that is tailored to the participant’s local usage or prices.
4
According to U.S. DOE (2014a), Table A4, space heating, space cooling, water heating, refrigeration,
clothes dryers, freezers, clothes washers, and dishwashers accounted for 62% of total residential energy
consumption in 2012. These end-uses represented in 2012 a total of 12.5 quadrillion Btu compared to 95.0
quadrillion Btu from all sectors and sources in 2012.
5
France mandated labels for a variety of appliances in 1976 and Japan, Canada, and the
United States followed soon after (Wiel and McMahon, 2001).5 The Energy Policy and
Conservation Act of 1975 mandated labels for certain appliances beginning in 1980.
Changes to the labeling program were made in the Energy Policy Act of 1992 which gave
rise to the EnergyGuide labels in their current form.
The Federal Trade Commission (FTC) is charged with enforcing these labeling
requirements. The FTC provides templates on its website for manufacturers to use and
the Energy Labeling Rule in the Code of Federal Regulations provides samples of
acceptable labels (Federal Trade Commission, 2014).
Information provision requirements for vehicles are similar. Since 1977, all new
cars and trucks sold in the United States must display information about vehicle fuel
efficiency. Until recently, labels reported estimated city, highway, and combined fuel
efficiency in miles-per-gallon (MPG). Starting with model year 2013, new labels provide
additional information including estimated gallons per 100 miles, annual fuel cost, and
five year fuel cost savings compared the average new vehicle. The inclusion of gallons per
100 miles brings the United States in line with the European Union which reports liters
per 100 kilometers.
Fuel economy labels on vehicles suffer from the same problem as do appliance
labels in using national energy prices to compute fuel savings and ignoring variation in
vehicle miles traveled across the states. Paradoxically, the improvement in fuel economy
labels on motor vehicles may exacerbate losses from inaccurate information on the labels.
When labels only reported miles per gallon, consumers had to undertake significant
mental computations to balance the cost savings from a more fuel efficient vehicle against
5
Wiel and McMahon (2001) discuss the early motivation for energy labels. Thorne and Egan (2002)
conduct qualitative interviews with focus groups about alternative graphical elements and other aspects of
EnergyGuide label design.
6
the higher purchase price (holding other attributes constant). The current labels now
report estimated five year cost savings for each vehicle relative to the fleet average. Now it
is more straightforward to balance cost savings from more efficient vehicles against
higher purchase price. But cost savings can differ significantly given differences in
average gasoline prices and driving patterns across states. Whether consumers will make
those mental adjustments is not clear.
C.
Focus on Air Conditioning
More accurate labels could be important for many different appliance types but in
our experiment we focus specifically on room air conditioners. More than 25 million
American households own one or more room air conditioners so this is a source of
energy consumption that is of large intrinsic interest.6 It is also a particularly lucid
example of an energy-efficiency investment for which consumers face a clear tradeoff
between purchase price and energy costs, and for which operating costs vary substantially
across states. Moreover, most consumers install room air conditioners themselves,
thereby avoiding any principal-agent problem that arises when contractors are involved
in selecting and installing equipment.
More broadly, residential air conditioning is of large and growing policy interest
nationwide because of the high level of energy consumption associated with it. In the
United States, there is air-conditioning in nearly 100 million homes (87% of homes), and
households spend an estimated $22 billion dollars annually on electricity for airconditioning.7 Table 1 shows that air conditioner usage is pervasive in all parts of the
6
U.S. Department of Energy, Residential Energy Consumption Survey 2009. See Table HC7.1 “Air
Conditioning in U.S. Homes”.
7
Data from U.S. Department of Energy (2009). See Table HC7.1 “Air Conditioning in U.S. Homes” and
Table CE3.6 “Household Site End-Use Consumption in the U.S.”.
7
country. The lowest share is in the West where one-third of households have no form of
air conditioning. The table also illustrates considerable variation in the shares of central
versus room air conditioning among those households with air conditioning with central
air conditioning dominating in all regions except the Northeast.
Figure 1 shows annual cooling hours by state from U.S. Department of Energy
(2014).8 This is the number of hours per year for which a household should expect to use
an air conditioner. On average, Americans face 1,265 cooling hours per year, but there is
enormous geographic variation. Within the continental United States average annual
cooling hours range from 310 in Maine to 2,771 in Florida, almost a 9:1 ratio.
Figure 2 shows average residential electricity prices by state for 2012 from U.S.
Department of Energy (2013a), Table 2.10. The average price is 12.4 cents per kilowatt
hour, but again there is substantial geographic variation. The lowest electricity prices in
2012 were found in Louisiana (8.4 cents), while New York had the most expensive prices
(17.6 cents), so more than a 2:1 ratio. Figure 2 is only showing variation in prices across
states. But there is variation within states across utilities as well. Using data from the 2013
EIA Form 861, we computed the standard deviation of residential electricity prices by
utility across the United States. The standard deviation in prices across the country is 3.7
cents per kWh. The standard deviation across states is 5.22 cents per kWh while the
standard deviation within states is only 2.7 cents per kWh. This much lower variation
within states suggests the potential for improving information with state-specific labels.
Annual operating cost for a room air conditioner depends on cooling hours and
electricity prices according to this simple equation,
8
U.S. Department of Energy (2014) reports annual cooling hours for room air conditioners for 218 U.S.
cities. We aggregated to the state level taking a weighted average of cities within each state weighting by
population.
8
(1)
Annual
Operating Cost
(dollars)
=
Annual
Cooling
Hours
*
Electricity Price
(dollars
per watt hour)
*
Size of Air
Conditioner,
(BTUs)
/
Energy-Efficiency
Ratio of Air
Conditioner
(BTUs per watt)
The “energy-efficiency ratio”, or EER, of an air conditioner is the ratio of the unit’s
cooling capacity (in BTUs) to its electricity consumption (in watts). The higher the EER,
the more energy-efficient the air conditioner. Figure 3 shows annual operating costs for a
medium-sized (10,000 Btu), medium-efficiency (10.0 EER) room air conditioner by state.
Operating costs vary widely across states, from $28 per year in Washington to $316 per
year in Florida, more than an 11:1 ratio. The geographic pattern reflects variation in both
cooling hours and electricity prices.
3.
Experimental Design
A.
Overview
Our experiment was implemented through Time-Sharing Experiments for the
Social Sciences (TESS), an NSF-funded program aimed at making it easier for academics
to run online experiments. TESS contracts with GfK (formerly “Knowledge Networks”) a
company that administers surveys and experiments using a nationally-representative
panel which they call the KnowledgePanel. This platform has been widely used by
economists, see, e.g., Allcott (2013), Allcott and Taubinsky (Forthcoming), and Newell
and Siikamaki (2014).
The KnowledgePanel is a nationally representative panel of some 55,000 adults
selected using random-digit dialing and address-based sampling (GfK, 2013). Participants
are provided with a computer and free internet service if they do not already have it.
From this panel, GfK constructs samples to respond to surveys and participate in
experiments on a wide variety of topics. Samples are constructed to represent the
9
underlying population of interest and upon completion of the survey or experiment,
study-specific sample weights are provided to ensure that the observable characteristics of
the final sample match the characteristics of the population of interest (GfK, nd). The
TESS-funded surveys put limits on sample size and the number of questions. For our
experiment, GfK asked 3,744 participants to take the survey, of whom 2,440 completed
the experiment (completion rate of 62.5 percent).
Participants in our experiment were asked to make three hypothetical purchase
decisions. Each decision involved selecting one of three room air conditioners that varied
by purchase price and expected annual energy cost. Participants were told that the three
air conditioners were otherwise identical except for these features. And, as we explain in
the appendix, we designed the choice sets carefully to maximize the precision of our
estimates.
We designed the experiment as a simple randomized controlled trial with
participants randomly assigned to either the control group or the treatment group.
During the experiment, the only difference between these two groups was the labels
which they were shown. The control group was shown the current EnergyGuide labels
which report operating costs based on national average electricity prices and typical
national usage. The treatment group, in contrast, was shown labels which report
operating cost based on average electricity prices and usage for the state in which each
participant resides. Finally, at the end we asked a short set of questions to elicit how well
the participants understood the labels they had just seen and to assess their knowledge
about state and national electricity prices and air conditioner usage. GfK also provided us
with a rich array of socio-economic information about the participants collected from
previous surveys. See the appendix for the complete survey instrument and list of
variables.
10
B.
The Treatment
Figure 4 shows examples of the labels we showed participants in the experiment.
Participants in the control group saw labels like the one on the left. This is the current
EnergyGuide label, and it shows estimated yearly energy cost based on national average
electricity costs and usage.9 Participants in the treatment group saw labels like the one on
the right. This particular label is for a participant in Iowa. The estimated yearly energy
cost is calculated based on the average residential price of electricity in Iowa ($0.1082 per
kWh) and the average usage in Iowa (828 hours per year). These state-specific labels were
tailored to the state of residence of each participant.10 That is, participants in the
treatment group from Iowa saw the Iowa label, and participants in the treatment group
from Nevada saw the Nevada label. Moreover, for all state-specific labels, we adjusted the
cost range to reflect the relevant range for that particular state. Because energy costs scale
linearly, this meant that the slider bar and “triangle” were positioned in the same place in
control and treatment labels.
In all cases, our labels are for a medium-sized (10,000 Btu) window unit. In
addition to reporting the estimated yearly energy cost in dollars, the label also reports the
unit’s EER, and further below, the label includes the language “Your cost will depend on
your utility rates and use.” Finally, the bottom of the label provides three bullets with
additional details. The first bullet explains that the cost range is based only on models
with similar capacity and characteristics. The second bullet explains how the energy cost
was calculated. This is important for our experiment, and we varied the text here
9
The actual EnergyGuide labels for room air conditioners report estimated annual energy cost based on
750 hours of usage. This has long been used as a rule-of-thumb, for example by the Association of Home
Appliance Manufacturers, but average usage in the United States is actually significantly higher. We use
1265 hours of usage per year based on the data that we use to calculate state-specific energy costs from U.S.
Department of Energy (2014). In all other ways, our labels are identical to the current EnergyGuide labels.
10
The KnowledgePanel programmers programmed the experiment so that the appropriate state-specific
label was automatically shown to each participant in the treatment group in a seamless fashion so that the
survey experience was identical across the control and treatment groups.
11
depending on treatment status. For the control group, the text reads, “Estimated energy
cost based on a national average electricity cost of 12.4 cents per kWh and national
average usage.” For the treatment group, the text reads, "Estimated energy cost based on
average electricity costs and usage for [State Name]." Finally, the last bullet points
consumers to the FTC website for more information.
C.
Balance in Sample
Before moving on to results, we test for balance between the control and
treatment groups. Since treatment status was randomly assigned, we expect very similar
characteristics in the two groups. Table 2 reports mean characteristics for the control and
treatment groups as well as p-values from tests that the means are equal. We report
weighted means using the sampling weights that were constructed specifically for our
experiment. This socio-economic information including political party affiliation was
collected from the individuals in the KnowledgePanel by GfK during previous surveys.11
Not surprisingly, given the design of the experiment, we fail to reject equality of
means between the two groups for any of the socioeconomic characteristics. The p-values
of 1.0 for educational status, sex, and race reflect the fact that the experiment-specific
sampling weights are balancing on these attributes.12 The mean characteristics also match
national data quite well. For example, the proportion of households with central air
conditioners (65.5 and 67.5 percent) is similar to the national average from the 2009
11
Political party affiliation is measured by GfK as "strong", "not strong", or "leans." We constructed
indicator variables for Democratic and Republican affiliation based on whether each participant indicated
"strong" or "not strong" support for a particular party.
12
The unweighted means are also very similar between the control and treatment groups. We also
computed p-values for equality of means between the two groups with the unweighted data and we
continue to find p-values in excess of 10 percent for the demographic and economic characteristics. In
addition we ran a weighted regression of a treatment indicator variable on all the variables in Table 2. The
F-statistic for the joint test that all the estimated coefficients are zero has a p-value of 0.75.
12
Residential Energy Consumption Survey reported in Table 1 (63 percent). The fraction of
participants with high school and college degrees is also similar to data from the U.S.
Census Bureau.
Despite households being randomly assigned to control and treatment groups, the
average residential electricity price is slightly higher in the control group and statistically
significant at the 10 percent level. Consequently, average yearly energy costs are also
slightly higher in the control group, though this difference is not statistically significant.
We attribute these modest differences to sampling variation and in our preferred
estimates will control for state fixed effects.
4.
Results
We present results in this section as follows. First we provide a simple graphical
depiction of our main results. We then turn to a regression framework to quantify the
magnitude of the effect controlling for state-fixed effects and other observable
characteristics, and we compare treatment effects across subsets of participants. Finally,
we use our preferred estimates to calculate aggregate national impacts.
A.
Graphical Evidence
As a first cut at the data, we compare the average characteristics of the air
conditioners selected by the treatment and control groups. We hypothesize, for example,
that participants living in states with high electricity prices will respond to more accurate
labels by choosing more energy-efficient air conditioners (i.e. with a higher EER). The
same prediction can be made for participants living in states with a large number of
annual cooling hours.
13
Figure 5 provides an initial attempt to answer our central research question. We
divided states into those with low, medium, and high operating costs. Specifically, we
ranked states by estimated annual energy cost (average state electricity price multiplied by
average state usage) and assigned states to these three categories based on whether the
state was in the lower, middle, or upper third of all states. For each group of states, we
plot the mean energy-efficiency of air conditioners selected by the treatment and control
groups. In addition to plotting these means, the figure also includes 95 percent
confidence intervals for each group constructed using standard errors clustered by
participant.
The results are striking. The participants who see the current EnergyGuide labels
choose similar levels of energy-efficiency in all three groups of states. This is interesting
and perhaps surprising given the enormous variation in cooling hours and electricity
prices across states that we documented earlier. The participants who see state-specific
labels choose less energy-efficient air conditioners in low-cost states and more energyefficient air conditioners in high-cost states. This suggests a more efficient allocation of
energy-efficiency. The returns to energy-efficiency are higher in states with high
operating costs because electricity expenditures are a larger share of the total cost of
cooling.
While illustrative, this figure does not control for electricity prices and other
factors that are imperfectly balanced between the treatment and control groups. Nor does
it allow us to quantify the cost of any misallocation of energy efficiency across
households. We turn to that analysis next.
B.
Measuring the Lifetime Cost of Appliance Ownership
With energy-efficiency investments the relevant measure is the lifetime cost of the
14
appliance. Lifetime cost (LTC) is the sum of an appliance’s purchase price (PP) and the
present discounted value of its annual energy costs (EC) over the appliance's lifetime.
Specifically
1
(2)
1
where ρ is the consumer's discount rate and T is the expected operating life of the
appliance.13
Our conjecture is that the group shown state-specific labels will make better
choices leading to lower average lifetime cost.14 When we make these calculations we use
a twelve-year appliance lifetime and use a discount rate which we estimate from our
data.15 Given the considerable discussion in the energy literature on the relevant discount
rate for thinking about energy-efficient capital, we also report results based on other
discount rates. But as a starting point, we believe it is reasonable to estimate a discount
rate using our data following long standing practice in the literature. Specifically, we first
analyze the data using a discrete choice model as has been done in previous studies of
consumer take-up of energy efficient appliances.16 This allows us to estimate an average
discount rate for the sample, a necessary input for calculating expected lifetime appliance
cost using equation (2) above.
13
We assume that the best estimate of future electricity prices is the electricity price at time of purchase.
This is consistent with estimates from U.S. Energy Information Administration (2014a) which shows a flat
ten year real price trend for predicted retail electricity prices.
14
Lifetime cost is an appropriate measure of welfare in our context because the air conditioners in our
experiment are otherwise undifferentiated. With actual air conditioners, consumers also derive utility from
the manufacturer brand, color, ease of use, etc. The difference is that these other characteristics are easily
observable so appliance buyers are already making efficient purchase decisions along these margins, and we
would not expect those choices to change materially with changes in EnergyGuide label design.
15
The U.S. Energy Information Administration (2014b) assumes room air conditioners have a minimum
life of 8 years and a maximum life of 16 years. EIA assumes an approximately linear retirement schedule so
the average expected lifetime is 12 years.
16
Hausman (1979) and Dubin and McFadden (1984) are seminal papers in this literature.
15
Participants are assumed to choose the appliance that yields the highest level of
utility,
,
(3)
where i indexes the participant and j indexes the different air conditioner alternatives.
Purchase prices
annual energy costs
are the same for all participants regardless of where they live, but
vary across participants.17 The idiosyncratic term
is assumed
to be independent across participants and alternatives and have an extreme value
distribution so the choice probabilities take the well-known conditional logit form.
Table 3 reports estimates and standard errors. Both coefficient estimates are
negative as expected. The ratio of the coefficient estimates on purchase price and energy
cost is 0.174, indicating that participants are willing to tradeoff $0.17 in purchase price for
a $1.00 change in annual energy costs. This corresponds to a discount rate (ρ) of 13.7
percent assuming a 12-year lifetime.18 In the results which follow we report lifetime costs
using this discount rate as well as alternative discount rates corresponding to a ratio of
coefficients that are 5 percentage points higher and lower. As will become clear, our
qualitative results are not affected by the discount rate we choose, but the magnitude of
the measured cost savings from state-specific labels is sensitive to the discount rate.
C.
Regression Estimates
We estimate regressions of the following form,
17
In particular, we assume that participants make decisions based on the information provided on the label.
For the control group, this is based on national average electricity prices and usage, and for the treatment
group, this is based on their state’s electricity prices and usage. We have also estimated the model restricting
the sample to include the treatment group only, and our estimate of the discount rate is similar.
18
This is similar to recent estimates in the literature from vehicle purchases including Busse, Knittel and
Zettelmeyer (2013) and Allcott and Wozny (2014).
16
∙
(4)
where the dependent variable
′
is one of our three different measures of cost (purchase
price, annual energy cost, or lifetime cost) based on the purchase decisions made by the
participants. The subscript indexes participant i, purchase decision j (j = 1, 2, 3), and state
s. Energy costs were calculated for all participants using state-specific measures of cooling
hours and electricity prices, and thus reflect our best estimate of actual operating costs
regardless of which labels the participant was shown.19 Regressions are estimated using all
7,275 choices made by the 2,440 participants in our online experiment. We estimate these
models in levels, but we have also estimated specifications in which costs are measured in
logs and the results are similar.
The covariate of interest is
, an indicator variable equal to 0 if the
individual is in the control group and 1 if in the treatment group. Thus, the treatment
effect
is the estimated difference in cost between the treatment and control groups, after
controlling for covariates. The vector X includes household income and indicator
variables for college graduate, non-white, married, age 65 and older, and political
affiliation. We also control for state fixed effects ( ). These controls increase the
precision of our estimates and correct for the modest imbalance in observed
characteristics between the treatment and control groups observed in Table 2.
Identification of
comes from within-state comparisons between participants in the
treatment and control groups.
Table 4 reports the regression estimates. The treatment group paid on average
19
These calculations implicitly assume that the price elasticity of demand for cooling is zero (i.e. that there
is no “rebound” effect). A richer framework would describe air conditioning as a household production
problem in which thermal comfort is traded off against electricity expenditure. Allowing for a non-zero
elasticity would increase the lifetime pecuniary cost of an energy-efficient unit, but also provide utility in
the form of improved thermal comfort. Because households are choosing usage levels optimally, these two
components will be similar in magnitude for small differences in energy-efficiency.
17
$3.44 more in purchase price than the control group, indicating slightly more investment
in energy-efficiency. We hypothesized that the state-specific labels would improve the
allocation of energy-efficiency investments across households, but there was no clear
prediction for purchase prices so this is not particularly surprising. Annual energy cost is
$2.36 lower on average in the treatment group and is statistically significant at the 10
percent level.
We are most interested in the impact on lifetime cost. The reduction in annual
energy cost accumulates over the lifetime of the air conditioner, resulting in significantly
lower lifetime costs from state-specific labels. On average, lifetime costs are $10.12 lower
in the treatment group than the control group. This estimate is statistically significant at
the 1% level. This reduction in lifetime costs is consistent with Figure 5 and indicates a
better allocation of energy-efficient air conditioners across states.20 Some of the other
coefficient estimates are also interesting. Lifetime cost is decreasing in household income
and education. Also, non-white participants pay considerably more in lifetime cost and
older participants pay considerably less. Finally, Democrats spend about the same
amount in lifetime cost, but there is suggestive evidence that Republicans spend
somewhat more.
It is worth noting that the fit of the model differs substantially across dependent
variables. In the first column the R2 is only 0.045, indicating that these decisions are
driven mostly by idiosyncratic factors. The R2 in the second column is much higher
(0.781) because the state fixed effects capture the variation in energy costs driven by
electricity prices and usage. And the R2 in the third column is the highest of all (0.916).
20
Results are similar in specifications where we control for whether each participant has central air
conditioning, room air conditioners, or no air conditioning. We also ran regressions on each sub-group
separately and find negative coefficients on the treatment variable in all three regressions, but only
statistically significant results for survey participants with central air conditioners. The lower statistical
significance reflects, in part, the smaller sample sizes. Less than one-third of the survey participants do not
have central air.
18
Lifetime costs are easier to predict because differences in purchase price offset differences
in the present discounted value of energy costs, so that the variation in lifetime cost has
more to do with cooling hours and electricity prices than with the energy-efficiency of the
selected appliances.
D.
The Allocation of Energy Efficiency across Regions
Table 5 reports additional regression estimates. Focusing on cost savings across
the entire sample masks important heterogeneity. As suggested by Figure 5, it may well be
that participants in low-cost states respond differently to state-specific labels than
participants in high-cost states. The top row corresponds exactly to the regression
estimates in Table 4, but also includes estimates of lifetime cost corresponding to
alternative values of the discount rate (ρ). Estimated savings increase to $15.60 with a 6.7
percent discount rate and fall to $7.09 with a 19.8 percent discount rate. In all cases, the
savings are statistically significant at the 5 percent level or lower.
For the regressions reported in the second through fourth rows, the sample is split
into three parts corresponding to low-, middle-, and high-energy cost states. As we saw
initially with Figure 5, the impact of state-specific labels varies considerably across
groups. Participants in low-cost states spend less upfront on air conditioners, and incur
less overall lifetime cost. With a 13.7% discount rate, lifetime savings are $6.78, a
difference that is statistically significant at the 5 percent level. Participants in mediumcost states incur about the same amount in overall lifetime cost. For these states, statespecific labels provide information that is very similar to the current EnergyGuide labels,
so it makes sense that there would not be large differences in behavior. Finally,
participants in high-cost states spend considerably more upfront on air conditioners, and
then incur considerably lower lifetime costs, ranging from $12.81 to $41.61 for the
19
discount rates we consider. In all cases the lifetime savings for this group are statistically
significant at the 5 percent level.
E.
Aggregate Savings Nationwide from State-Specific Labels
Households can make two kinds of mistakes when buying air conditioners with
inaccurate information about operating costs. Households in low-cost states (e.g.
Massachusetts) may purchase overly energy-efficient air conditioners despite the fact they
will operate these air conditioners only a few days a year. In our experiment, participants
from low-cost states save nearly $7 on average in lifetime costs with better information.
Conversely, households in high-cost states (e.g. Florida) may purchase less energyefficient air conditioners than is optimal given the expected heavy usage in that state. In
our experiment, participants from high-cost states save $23 on average in lifetime costs
with better information. Overall, better information leads to private gains of over $10 per
air conditioner purchase.
Table 6 reports the aggregate national savings implied by our estimates. That is,
the table reports how much consumers would save nationwide from a shift to statespecific EnergyGuide labels. At a per-unit savings of $10.12 and nationwide annual sales
of 4.4 million units, the cost savings for room air conditioners sold in a given year is $44.5
million.21 Discounting future year savings at 13.7 percent (and assuming no increase in
sales or annual energy costs), we get a present discounted value of savings of $370 million.
More broadly, our findings suggest that state-specific labels would improve
21
These calculations ignore potential responses by appliance manufacturers and retailers. In the short-run,
firms might adjust pricing in response to the change in demand for different models. The U.S. appliance
market has become more competitive with the recent entry of LG, Samsung, and other international
manufacturers, but firms are still able to charge significant markups particularly for high-end models
(Houde (2014a); Spurlock (2014)). Moreover, in the long-run manufacturers might respond to better
information by changing the set of appliances offered for sale.
20
purchase decisions not just for room air conditioners, but also for many different types of
appliances. Central air conditioners, furnaces, and heat pumps are obvious examples
because cooling and heating demand varies across states. But appliances like refrigerators,
freezers, clothes washers, and dishwashers could also benefit from state-specific labels. As
we showed earlier, residential electricity prices vary by more than 2:1 across states, so
there are significant potential efficiency gains from improved information even for
products with little predictable cross-state variation in usage.22
These benefits can be compared to the cost of implementing state-specific labels.
Requiring manufacturers to ship appliances with state-specific labels would not require
any additional appliance testing. The FTC currently maintains label templates that
manufacturers can download. Instead of one template per appliance, the FTC would need
to maintain 50 different templates, one for each state, perhaps accessible through a dropdown menu. At the same time it might also make sense to automate the simple
calculation required to fill in estimated yearly energy cost. Although these changes with
the FTC website would presumably be relatively inexpensive, the more substantive
administrative burden would fall on the manufacturers themselves. The challenge for
manufacturers is that labels are often attached to appliances even before it is known
where they are going to be shipped. Moreover, appliances are frequently rerouted across
states. For example, an appliance originally intended for California can end up Nevada. It
might make sense to use region-specific labels, rather than state-specific, to reduce the
amount of relabeling that is required and/or to ship appliances with labels prepared for
several different states.23
22
While we have not addressed the issue of externalities associated with appliance use and the interaction
with better labels, we note that carbon pricing, for example, would change – and perhaps increase – the
regional variation in electricity prices. See, for example, Graff Ziven, Kotchen and Mansur (2014).
23
The U.S. Department of Energy has taken a region-based approach with new minimum efficiency
standards for air conditioners and heat pumps. The United States has been divided into three regions
21
An alternative deployment option would be add a QR scan code to existing labels
which consumers could scan with their smart phones.24 The phone would then
automatically display a label with state-specific or even county-specific annual energy
costs. This would require the FTC to maintain a website with data on average annual
energy costs that would be queried by the phone's QR scan app. The cost of including a
QR scan code on labels would be near zero, and the cost to the FTC of developing the
software and maintaining such a system would be relatively low, though whether or not
consumers would use the information is unclear. Another related deployment option
would be to develop an automated system for online retailers. By law retailers must make
EnergyGuide labels available for online shoppers and an automated system would display
labels that are tailored to each consumer’s state or county of residence. This
customization would be somewhat easier logistically than the physical labels because of
the issue of not knowing where appliances are going to be shipped.
5.
Underlying Mechanisms
Having documented substantial treatment effects from the introduction of state-
specific EnergyGuide labels, we next turn to an analysis of the underlying mechanisms
driving our results. Specifically, we ask three questions: (1) Do participants understand
the labels? (2) Do participants know whether their state's annual energy cost from
operating an air conditioner is higher or lower than the national average? (3) Do
participants take local factors into account when selecting a level of efficiency?
(North, Southwest, and Southeast) and, beginning January 1, 2015, air conditioners and heat pumps
manufactured for the two Southern regions must meet a higher minimum efficiency standard. See U.S.
Court of Appeals Case # 11-1485, April 24, 2014 for details.
24
The new EPA vehicle mileage labels that went into effect beginning with model year 2013 include a QR
scan code providing smart phone access to online information about fuel economy and environmental
factors.
22
A.
Do Participants Understand the Labels?
Table 7 shows the responses to two multiple choice questions we asked
participants immediately after they made their hypothetical appliance choices. The exact
wording of the questions is provided in the table. These questions were aimed at
investigating how well participants understood the labels they had just seen. Participants
were not able to go back and look again at the labels before answering the questions.
Overall, participants demonstrate a remarkably poor level of label comprehension.
Over half the participants were not sure whether the national or state electricity price was
used to compute yearly costs. And among those who had an opinion, many incorrectly
answered the question.26 There is no statistical difference between the percentage of each
group that thought it was the national average price (33.6 versus 30.8 percent). However,
the treatment group was more likely to answer correctly that it was the state price (17.0
versus 10.1 percent). This difference is highly statistically significant, but indicates that
only a relatively small fraction of participants in the treatment group actually realized
they were seeing operating costs calculated using state-specific information. The
responses are similar for the question about what usage level was used. Again, over half of
the participants were not sure whether national or state information was used. And again,
among those who expressed an opinion there is a large fraction of incorrect responses.
B.
Do Participants Know How Their State Compares?
Part of the rationale for the current EnergyGuide labels is that individuals should
be able to “translate” the operating cost information to incorporate information about
local electricity prices and usage. The labels include the phrase, “Your cost will depend on
26
Allcott (2011a) provides related evidence from vehicle purchase decisions. When purchasing a vehicle,
40% of respondents report not thinking “at all” about fuel costs, and an additional 35% report thinking
“some” about fuel costs but not making any calculations.
23
your utility rates and use.” And, at least in theory, an individual could transform the
estimated yearly energy cost to a more meaningful measure reflecting local information.
This hinges, however, on individuals having some sense of how their local energy prices
and usage compare to the national average.
Table 8 shows the responses to two multiple choice questions aimed at evaluating
this knowledge. We first asked participants how electricity prices in their state compare to
the national average. More than two-thirds of the participants answered that they were
not sure and, overall, only 20% of participants were able to correctly answer the question.
Participants have a somewhat better understanding of how their air conditioning usage
compares to the national average. A larger fraction of participants felt confident in taking
a position (60 percent versus 30 percent) and, overall, 40% of participants were able to
correctly answer the question.27
C.
Do Participants Take Local Factors Into Account?
The evidence from the previous subsections suggests that consumers are not
going to be able to mentally adjust the information in the current EnergyGuide labels to
account for local factors. Participants overall do not fully understand the information
they are being shown, nor do they consistently know how electricity prices and usage in
their state compare to the national average. In this section we formalize this conjecture by
testing whether state-level electricity prices and usage have any predictive power for
purchase decisions.
Table 9 shows regression estimates from two separate regressions. The dependent
27
We also examined responses separately for the treatment and control groups and the distribution of
responses is very similar and not statistically different (p-values 0.41 and 0.70). This suggests that
participants in the treatment group are not inferring anything about their state’s electricity prices or usage
based on the labels they are shown.
24
variable in both regressions is the energy-efficiency of the selected air conditioner
(measured in EER). For the control group, neither the electricity price nor usage has a
statistically-significant effect on energy-efficiency. The p-value for the joint null
hypothesis of no influence is 0.24. Moreover, the sign of the estimated coefficient on
price is negative, counter to what theory would suggest. This is pretty surprising and
provides no evidence that participants in the control group are able to mentally adjust the
information provided in the labels to account for local operating costs.
In contrast, for the treatment group, both price and usage are positive and jointly
strongly statistically significant. While we cannot reject the null that the coefficient on
price is zero at the 5 percent level, it is statistically significant at the 10 percent level and
the coefficient on usage is significant at the 1 percent level. This is what we would expect
given the savings documented in Tables 4 and 5. Importantly this evidence does not
support the hypothesis that participants are making any mental adjustments. It seems
more likely, given the lack of label comprehension in Table 7 that participants are simply
responding to the operating cost presented in the state-specific labels.
D.
Complementary Revealed Preference Evidence
An important question is how any of these results would generalize to actual
choices. With good reason, economists have long been skeptical about interpreting results
from stated-choice experiments (Hausman, 2012). Without any real “skin in the game,”
it is not at all clear that participants in an online experiment are going to make the same
choices that they would when faced with real financial consequences. We have attempted
to reduce these concerns by focusing on a concrete purchase decision that is designed to
look similar to actual decisions that individuals face, but we recognize the limitations
inherent with stated choice.
25
In our context, it is not even possible to make strong statements about the
direction of bias. On the one hand, better labels might tend to be less effective than in the
real-world because there is no actual money at stake, so participants are going to tend to
answer these questions quickly and perhaps not read the fine print. On the other hand,
our stated-choice setting removes some additional factors like appliance manufacturer
and differences in sizes, color, and other design considerations potentially leading
participants to focus more on these labels than they would in the real-world. It is
impossible to know which of these potential biases is more important.
Federal law requires that EnergyGuide labels be displayed on all major appliances
sold in the United States. Thus, it is not at all straightforward how to replicate this online
experiment in the field. Strictly speaking, it would be illegal to go into an appliance
retailer and replace the current labels with labels providing state-specific information.
One possibility would be to supplement the existing labels with additional information of
some form. Although this would indeed be interesting, the results of such an experiment
would be somewhat difficult to interpret. Such a treatment would inevitably increase
attention on operating costs, and it would be difficult to disentangle the impact of that
attention from the pure information content.
Another approach to validating our stated-choice experiment is to look for
complementary evidence from actual choices. Figure 6 shows the fraction of new central
air conditioners sold in each state in 2009 that has an Energy Star rating.28 What is
striking about this figure is the lack of correlation between these choices and the pattern
of operating costs we showed in Figure 3. Operating costs are highest throughout the
28
In order to obtain an Energy Star rating (and so display the Energy Star logo) product manufacturers
must meet design criteria that contribute to significant energy savings without sacrificing product quality,
features, and performance. Unit shipment data are from U.S. Department of Energy (2010) and come from
sales data that represent roughly 60 percent of the retail market. The shares in Figure 6 should therefore be
seen as indicative only.
26
South, from Texas through Louisiana, Mississippi, Alabama, Georgia, South Carolina and
Florida. So if choices are being made efficiently, we would expect to see large investments
in energy-efficiency in these states. Instead, the states with the highest Energy Star shares
are in the Northeast and upper Midwest. The cross-state correlation between the Energy
Star share and state average residential electricity prices is 0.41. But the correlation with
average usage is -0.33 and the correlation with estimated annual energy cost (electricity
price multiplied by usage) is -0.23.29 Thus, purchases of Energy Star air conditioners
appear to be systematically biased away from what would be required for efficiency.
The high penetration of Energy Star air conditioners in states like Vermont and
Massachusetts suggests that other factors including political ideology may come into play
when households make choices about energy-efficiency. Our experiment provides some
supportive evidence for this hypothesis. In particular, political party affiliation did seem
to matter for air conditioner choices in Table 4. While being affiliated with the
Democratic Party does not have a statistically significant effect, participants who are
affiliated with the Republican Party tend to choose less expensive (i.e. less energyefficient) air conditioners and thus spend more in annual operating cost.30 As always,
however, it is important to interpret these cross-sectional comparisons with caution.
Political ideology is not the only factor that could explain this geographic pattern of
Energy Star adoption. Air conditioning is less common in the North, so it tends to be
higher-income households making these purchases, and this compositional effect could
29
In related work, Jacobsen (2014) finds using panel data no evidence that electricity prices increase
purchases of Energy Star appliances, and Houde (2014b) finds using transaction-level data from a major
retailer relatively little sensitivity of appliance choices to local electricity prices.
30
Previous papers have documented similar correlations between political ideology and adoption of
energy-efficient vehicles and buildings (Kahn and Vaughn, 2009). One of the potential explanations that
has been suggested is that in "green" communities, driving an energy-efficient vehicle or owning an energyefficient building could be perceived as a symbol of "status" (Kahn, 2007). We are not aware of previous
attempts to correlate political ideology with air conditioner choices, but these purchases are considerably
less visible than vehicles and buildings, suggesting that other more intrinsic explanations may play a role.
27
provide an alternative explanation.
Nonetheless, this pronounced lack of correlation between operating costs and
choices provides some real-world corroboration for the evidence from our stated-choice
experiment. Revealed preference cannot tell us how much choices would be improved by
better information, but it does provide suggestive evidence that the current labels are not
working as well as they could. It may not be enough to simply say, as the current labels
do, that “Your cost will depend on your utility rates and use.” We may need to provide
better information to help consumers connect the dots.
E.
Discussion and Implications
The state-specific labels changed participants’ behavior, so participants are not
ignoring these labels completely. But at the same time, participants are not exerting the
effort that would be required to understand the information beyond a superficial level. In
the labels the annual operating cost appears in 24-point font, bigger than all other text.
Participants in the experiment appear to have read and internalized that one number, but
then failed to read or internalize anything else. Moreover, there is no evidence of
individuals spontaneously incorporating local information when they see only nationalaverage information.
Most participants do not make intertemporal decisions like this regularly. Getting
a decision like this exactly right would require real time and cognitive effort, so it makes
sense that participants may try to simplify these decisions, either consciously or
unconsciously. One way to simplify the problem is to take the headline operating cost
number as given, and ignore everything else. Whether this inattention is rational or
irrational is unclear. It could be that participants are weighing the potential benefits of
becoming perfectly informed against attention and other costs and choosing consciously
28
to be inattentive (Sallee, 2014). Or it could be that they have unconsciously switched into
an inattentive mode and could switch back at relatively low cost.
Another point that emerges from this analysis is the distinction between
information programs and energy conservation programs. While providing state-specific
information to households appears to lead to more economically efficient appliance
purchases, it does not necessarily mean that aggregate energy use will fall. Table 10 shows
that, in our experiment, electricity consumption, in fact, does go down, by an average of
16.5 kilowatt hours per year, driven by significant decreases in consumption high-cost
states. However, this need not be the case. In general, providing better information leads
energy consumption to decrease in high-cost states but increase in low-cost states.
Whether the net change in consumption is positive or negative depends on the type of
information provided and characteristics of the households receiving that better
information. But – and this is important – better information is efficiency enhancing
regardless of the effect on energy use.31
6.
Conclusion
Energy efficiency is critically important both as an element of a portfolio of
measures to reduce greenhouse gas emissions to address global climate change
(Intergovernmental Panel on Climate Change, 2014) as well as concerns about local
pollutants from the burning of fossil fuels. This paper contributes to our understanding of
the role information plays in shaping consumer purchase decisions as well as possible
instruments to improve purchase decisions for optimal levels of energy-efficient capital.
We find that better labels lead to better choices. State-specific labels decrease the
31
This ignores the fact that the private cost of energy may not match the social cost. For an in depth
analysis of the externalities associated with energy production and consumption, see National Research
Council (2009).
29
lifetime cost of air conditioning both in high- and low- operating costs states. In high-cost
states like Florida and Texas, consumers invest more in energy-efficiency and this
increase in upfront spending is outweighed by a substantial decrease in annual energy
expenditures. In low-cost states like Maine and Oregon, consumers invest less in energyefficiency and this decrease in upfront spending outweighs a modest increase in annual
energy expenditures.
Despite the improved allocation, there remains a puzzle: although participants
respond to the labels, they do so with a poor understanding about where this information
is coming from or what it means. Immediately after the experiment, most participants
cannot answer basic questions about the labels they just saw. One possible explanation
for the puzzle is that participants treat the label as WYSIATI. That is, when they look at
the labels they fixate on the main headline summary number in large font, while
essentially ignoring everything else. If this is correct, it has important implications for
label design. Most importantly, it suggests that we should be working hard to make sure
that the headline number is as accurate as possible, and that we should not assume that
households can “translate” information to reflect local or personal variation in prices and
usage. This conjecture suggests a fruitful line of future research, both in the lab and in the
field.
Our research has practical significance as well. The implied aggregate cost savings
from the improved allocation of appliances across households could be quite large
relative to the cost of implementing state-specific labels. Customized information could
improve decision making not only for air conditioners, but for many different types of
appliances. While the usage of most appliances does not vary geographically as much as
air conditioning, electricity prices vary by more than 2:1 across states, so there are
potentially significant efficiency gains from improved information even for products with
little variation in usage.
30
References
Allcott, Hunt. 2011a. "Consumers' Perceptions and Misperceptions of Energy Costs."
American Economic Review Papers & Proceedings, 101(3), 98-104.
Allcott, Hunt. 2011b. "Social Norms and Energy Conservation." Journal of Public
Economics, 95(9-10), 1082-1095.
Allcott, Hunt. 2013. "The Welfare Effects of Misperceived Product Costs: Data and
Calibrations from the Automobile Market." American Economic Journal: Economic Policy,
5(3), 30-66.
Allcott, Hunt and Todd Rogers. Forthcoming. "The Short-Run and Long-Run Effects of
Behavioral Interventions: Experimental Evidence from Energy Conservation." American
Economic Review.
Allcott, Hunt and Richard Sweeney. 2015. "Can Retailers Inform Consumers About
Energy Costs? Evidence from a Field Experiment."
Allcott, Hunt and Dmitry Taubinsky. Forthcoming. "The Lightbulb Paradox: Evidence
from Two Randomized Experiments." American Economic Review.
Allcott, Hunt and Nathan Wozny. 2014. "Gasoline Prices, Fuel Economy, and the
Energy Paradox." Review of Economics and Statistics, 96(10), 779-795.
Ayres, Ian; Sophie Raseman and Alice Shih. 2009. "Evidence from Two Large Field
Experiments That Peer Comparison Feedback Can Reduce Residential Energy Usage,"
National Bureau of Economic Research: Cambridge, MA, NBER Working Paper No.
15386.
Bertrand, Marianne and Adair Morse. 2011. "Information Disclosure, Cognitive Biases,
and Payday Borrowing." Journal of Finance, 66(6), 1865-1893.
Brounen, Dirk and Nils Kok. 2011. "On the Economics of Energy Labels in the Housing
Market." Journal of Environmental Economics and Management, 62(2), 166-179.
Busse, Megan R.; Christopher R. Knittel and Florian Zettelmeyer. 2013. "Are
Consumers Myopic? Evidence from New and Used Car Purchases." American Economic
Review, 103(1), 220-256.
Camilleri, Adrian R. and Richard P. Larrick. 2014. "Metric and Scale Design as Choice
Architecture Tools." Journal of Public Policy and Marketing, 33(1), 108-125.
Collaborative Labeling and Appliance Standards Program (CLASP). 2014. "Global S&L
Database," http://www.clasponline.org/en/Tools/Tools/SL_Search.aspx, Accessed on Sept.
25, 2014.
31
Davis, Lucas W. 2008. "Durable Goods and Residential Demand for Energy and Water:
Evidence from a Field Trial." RAND Journal of Economics, 39(2), 530-546.
Dubin, Jeffrey A. and Daniel L. McFadden. 1984. "An Econometric Analysis of
Residential Electric Appliance Holdings and Consumption." Econometrica, 52(2), 345362.
Federal Trade Commission. 2014. "Energyguide Labeling: Faqs for Appliance
Manufacturers," http://www.business.ftc.gov/documents/bus-82-energyguide-labels-faqs,
Accessed on October 9, 2014.
GfK. 2013. "Knowledgepanel Design Summary," http://www.gfk.com/Documents/GfKKnowledgePanel-Design-Summary.pdf, Accessed on September 26, 2014.
GfK. nd. "Gfk Methodology," provided with data; available from authors.
Gillingham, Kenneth; Richard G. Newell and Karen Palmer. 2009. "Energy Efficiency
Economics and Policy." Annual Review of Resource Economics, 1(1), 597-619.
Gillingham, Kenneth and Karen Palmer. 2014. "Bridging the Energy Efficiency Gap:
Policy Insights from Economic Theory and Empirical Evidence." Review of
Environmental Economics and Policy, 8(1), 18-38.
Graff Ziven, Joshua; Matthew J. Kotchen and Erin T. Mansur. 2014. "Spatial and
Temporal Heterogeneity of Marginal Emissions: Implications for Electric Cars and Other
Electricity-Shifting Policies." Journal of Economic Behavior and Organization, 107(Part A,
November), 248-268.
Hastings, Justine S. and Jeffrey M. Weinstein. 2008. "Information, School Choice, and
Academic Achievement: Evidence from Two Experiments." Quarterly Journal of
Economics, 123(4), 1373-1414.
Hausman, Jerry A. 1979. "Individual Discount Rates and the Purchase and Utilization of
Energy-Using Durables." Bell Journal of Economics, 10(1), 33-54.
Hausman, Jerry A. 2012. "Contingent Valuation: From Dubious to Hopeless." Journal of
Economic Perspectives, 26(4), 43-56.
Houde, Sebastien. 2014a. "Bunching with the Stars: How Firms Respond to
Environmental Certification," University of Maryland Working Paper: College Park, MD.
Houde, Sebastien. 2014b. "How Consumers Respond to Environmental Certification and
the Value of Energy Information," E2e Project: Berkeley, CA, Working Paper 007.
Hoxby, Caroline and Sarah Turner. 2013. "Expanding College Opportunities for HighAchieving, Low Income Students," Stanford, CA.
32
Intergovernmental Panel on Climate Change. 2014. "Climate Change 2014: Mitigation
of Climate Change, Working Group iii Contribution to the Fifth Assessment Report of
the Intergovernmental Panel on Climate Change,"
Jacobsen, Grant. 2014. "Do Energy Prices Influence Investment in Energy Efficiency?
Evidence from Energy Star Appliances," University of Oregon Working Paper.
Kahn, Matthew E. 2007. "Do Greens Drive Hummers? Environmental Ideology as a
Determinant of Consumer Choice." Journal of Environmental Economics and
Management, 54, 129-145.
Kahn, Matthew E. and Ryan K. Vaughn. 2009. "Green Market Geography: The Spatial
Clustering of Hybrid Vehicles and Leed Registered Buildings." B.E. Journal of Economic
Analysis & Policy, 9(2, Article 2), 1-24.
Kahneman, Daniel. 2011. "Thinking Fast and Slow," New York: Farrar, Strauss and
Giroux.
Kling, Jeffrey R.; Sendhil Mullainathan; Eldar Shafir; Lee C. Vermeulen and Marian
V. Wrobel. 2012. "Comparison Friction: Experimental Evidence from Medicare Drug
Plans." Quarterly Journal of Economics, 127(1), 199-235.
Metcalf, Gilbert E. 1994. "Economics and Rational Conservation Policy." Energy Policy,
22(10), 819-825.
Metcalf, Gilbert E. and Kevin A. Hassett. 1999. "Measuring the Energy Savings from
Home Improvement Investments: Evidence from Monthly Billing Data." Review of
Economics and Statistics, 81(3), 516-528.
National Research Council. 2009. Hidden Costs of Energy: Unpriced Consequences of
Energy Production and Use. Washington, DC: National Academies Press.
Newell, Richard G. and Juha V. Siikamaki. 2014. "Nudging Energy Efficiency Behavior:
The Role of Information Labels." Journal of the Association of Environmental and Natural
Resource Economists, 1(4), 555-598.
Revelt, David and Kenneth Train. 1998. "Mixed Logit with Repeated Choices:
Households' Choices of Appliance Efficiency Level." Review of Economics and Statistics,
80(4), 647-657.
Sallee, James M. 2014. "Rational Inattention and Energy Efficiency." Journal of Law and
Economics, 57(3), 781-820.
Spurlock, Anna. 2014. "Appliance Efficiency Standards and Price Discrimination,"
Lawrence Berkeley Lab: Berkeley, CA.
33
Thorne, Jennifer and Christine Egan. 2002. "An Evaluation of the Federal Trade
Commission's Energyguide Appliance Label," American Council for an Energy Efficient
Economy: Washington, DC.
U.S. Department of Energy. 2009. "Residential Energy Consumption Survey (RECS)
2009," http://www.eia.gov/consumption/residential/data/2009/, Accessed on September
25, 2014.
U.S. Department of Energy. 2010. "2009 Unit Shipment Data Collection Forms,"
https://www.energystar.gov/index.cfm?c=partners.unit_shipment_data_archives,
Accessed
U.S. Department of Energy. 2013a. "Electric Power Annual," Washington, DC,
DOE/EIA-0348.
U.S. Department of Energy. 2013b. "Energy Star Unit Shipment and Market Penetration
Report
Calendar
Year
2012
Summary,"
https://www.energystar.gov/index.cfm?c=partners.unit_shipment_data_archives,
Accessed on September 25, 2014.
U.S. Department of Energy. 2014. "Room Air Conditioner Savings Calculator,"
http://www.energystar.gov/buildings/sites/default/uploads/files/RoomAC_Calculator.xls,
Accessed on September 26, 2014.
U.S. Energy Information Administration. 2014a. "Annual Energy Outlook 2014,"
Washington, DC:
U.S. Energy Information Administration. 2014b. "Assumptions to the Annual Energy
Outlook 2014: Residential Demand Model," Washington, DC,
Wiel, Stephen and James E. McMahon. 2001. Energy-Efficiency Labels and Standards.
Washington, DC: Collaborative Labeling and Appliance Standards Program (CLASP).
34
Table 1. Air Conditioner Penetration in U.S. Homes
US
Northeast
Midwest
South
West
Central Air Conditioner
62%
35%
66%
82%
46%
Room Air Conditioner(s)
24%
50%
22%
15%
17%
Both Central and Room Air
Conditioners
1%
1%
2%
1%
1%
No Air Conditioner
13%
13%
9%
2%
36%
Note: This table describes air conditioner penetration in the United States by region as estimated in the
U.S. Energy Information Administration, Residential Energy Consumption Survey, 2009. We have
excluded a small share of households who report having central or room air conditioners but not using
them. Regions are defined using standard Census definitions as Northeast (CT, MA, ME, NH, NJ, NY, PA,
RI, VT), Midwest (IA, IL, IN, KS, MI, MN, MO, ND, NE, OH, SD, WI), South (AL, AR, DC, DE, FL, GA,
KY, LA, MD, MS, NC, OK, SC, TN, TX, VA, WV) and West (AK, AZ, CA, CO, HI, ID, MT, NM, NV,
OR, UT, WA, WY).
35
Table 2. Testing for Balance in Randomized Sample
Control Treatment p-value
(1)
(2)
(3)
Annual Household Income (in
72,817
70,848
0.363
dollars)
High School Graduate
0.874
0.874
1.000
College Graduate
0.289
0.289
1.000
Household Size
2.745
2.756
0.871
Married
0.533
0.533
0.981
Employed
0.582
0.564
0.408
Age 65 and older
0.174
0.179
0.723
Female
0.519
0.519
1.000
Nonwhite
0.338
0.338
1.000
Homeowner
0.728
0.695
0.100
Multiunit Property
0.250
0.256
0.718
Household has a Central Air
0.655
0.675
0.322
Conditioner
Democratic Affiliation
0.316
0.314
0.942
Republican Affiliation
0.217
0.244
0.115
Average Residential Electricity
Price in the State of Residence
12.49
12.32
0.088
(cents per kWh)
Average Annual Hours of Air
Conditioning Use in the State
1,260
1,265
0.840
of Residence
Annual Cost of Operating a
Medium-Sized Room Air
154.58
153.04
0.601
Conditioner in the State of
Residence (in dollars)
Note: This table tests for balance between the control and treatment groups.
There are 1231 participants in the control group and 1209 participants in the
treatment group. Columns 1 and 2 report means of the variables listed in the
row headings, weighted using sampling weights. Proportion high school
graduate, college graduate, employment status, and the other individual
characteristics correspond to the individual in each household who participates
in the KnowledgePanel, not for the head of household. The annual cost of
operating a medium-sized room air conditioner is calculated for a 10,000 Btu
unit with an EER of 10.0. Column 3 reports p-values from tests that the
weighted means in the two groups are equal.
36
Table 3. Conditional Logit Regression Results
Variable
Coefficient Estimate
Purchase Price (PP)
-0.00223
(0.0004)
Energy cost (EC)
-0.01281
(0.0016)
Ratio of the Coefficient
Estimates on PP and EC
0.174
(0.013)
Implied Discount Rate ( )
0.137
(0.017)
Note: This table reports coefficients from a conditional logit model
estimated using all 7,275 choices made by the 2,440 participants in
our online experiment. There are slightly less than 3 choices per
participant because a small number of participants failed to finish
the experiment. The implied discount rate is calculated using an
assumed 12-year appliance lifetime. Observations are weighted
using sampling weights. Standard errors, clustered by state, are
reported in parentheses. All coefficient estimates are statistically
significant at the 1 percent level.
37
Table 4. Cost Impacts of State-Specific Labels, Regression Estimates
(1)
(2)
(3)
Annual
Purchase Price
Lifetime Cost
VARIABLES
Energy Cost
Treatment
Household
Income (x1000)
College Graduate
Nonwhite
Married
Age 65 and Over
Democrat
Republican
Constant
Observations
R-squared
3.436
(4.996)
-2.357*
(1.344)
-10.123***
(3.765)
0.307***
(0.055)
1.771
(5.812)
-13.869**
(6.187)
16.511***
(5.321)
18.366***
(6.131)
0.000
(6.066)
-9.026
(6.371)
365.458***
(6.456)
-0.081***
(0.014)
-1.738
(1.532)
5.532***
(1.740)
-3.232**
(1.415)
-5.816***
(1.575)
-0.413
(1.657)
3.102*
(1.685)
155.148***
(1.686)
-0.161***
(0.040)
-8.224*
(4.226)
17.954***
(5.004)
-2.078
(3.933)
-15.087***
(4.453)
-2.375
(4.619)
8.817*
(4.843)
1,257.890***
(4.674)
7,275
0.045
7,275
0.781
7,275
0.916
Note: This table reports coefficient estimates and standard errors from three separate
least squares regressions. The dependent variable varies across regression as
indicated in the column headings. Lifetime cost is calculated using a discount rate ( )
of 13.7 percent. All regressions include state fixed effects in addition to the covariates
listed in the row headings. The sample includes all 7,275 choices made by the 2,440
participants in our online experiment. In all regressions observations are weighted
using sampling weights. Standard errors are clustered by participant.
*** p<0.01, ** p<0.05, * p<0.1
38
Table 5. Cost Impacts of State-Specific Labels, Additional Regression Estimates
Lifetime Cost
Purchase Price
Annual Energy
=.137
=.067
Cost
(1)
(2)
(3)
(4)
Entire Sample
$3.44
-$2.36*
-$10.12***
-$15.60**
(5.00)
(1.34)
(3.76)
(6.61)
Low Operating Cost States
-$12.49
$0.99
-$6.78**
-$4.48
(7.82)
(1.09)
(2.77)
(2.87)
Medium Operating Cost
$1.50
-$0.02
$1.37
$1.32
States
(9.35)
(2.31)
(4.40)
(9.58)
High Operating Cost States
$22.86**
-$7.98**
-$23.06**
-$41.61**
(9.02)
(3.23)
(9.98)
(17.39)
=.198
(5)
-$7.09***
(2.49)
-$8.06**
(3.58)
$1.40
(1.99)
-$12.81**
(5.97)
Note: This table reports coefficient estimates and standard errors corresponding to the treatment indicator variable from twenty separate
least squares regressions. Positive numbers indicate a higher price or cost for the treatment group. The dependent variable varies across
regressions as indicated in the column headings. Lifetime costs are calculated using the discount rates as indicated. All regressions include
state fixed effects as well as household income and indicator variables for college graduate, non-white, married, age 65 or over, and political
party affiliation. For the first row the sample includes all 7,275 choices made by the 2,440 participants in our online experiment. For the
regressions reported in the second through fourth rows, states are divided into three groups (terciles) based on average energy costs
(residential electricity prices multiplied by annual hours of air conditioning use) and then regressions are run using participants from each
subset of states. In all regressions observations are weighted using sampling weights. Standard errors are clustered by participant.
*** p<0.01, ** p<0.05, * p<0.1
39
Table 6. Implied Aggregate National Savings from State-Specific Labels
Lifetime Cost Savings per Room Air
$10.12
Conditioner
Annual U.S. Sales of Room Air Conditioners
4.4 million
Total Cost Savings per Year
$44.5 million
Total Cost Savings – All Future Years
$369.5 million
(discounted at 13.7 percent)
Note: This table reports the implied aggregate national savings implied by our estimates. Lifetime cost
savings per air conditioner come from the full-sample regression estimate corresponding to a discount rate
of 13.7%. Annual sales of room air conditioners come from U.S. Department of Energy (2013b). Total cost
savings for all room air conditioners is the product of the first and second rows. The final row reports the
present discounted value of total cost savings implied by a permanent switch to state-specific labels.
40
Table 7. Testing Label Comprehension
Participants Shown
Participants
p-Values for
Current Labels
Shown
Equality of
(i.e. Control Group)
State-Specific
Proportions
Labels
Across Groups
(i.e. Treatment
Group)
(1)
(2)
(3)
What electricity price was used to calculate estimated yearly energy cost in the Energy Guide
labels you were shown?
The average electricity price in the
30.8%
0.152
33.6%
United States
The average electricity price in my
10.1%
0.000
17.0%
state.
I’m not sure.
56.3%
52.2%
Operating costs for an air conditioner depend on the cost of electricity and the number of hours
the air-conditioner is used. What usage level was used to calculate estimated yearly energy cost in
the Energy Guide labels you were shown?
The average usage level for air
32.2%
0.392
33.9%
conditioners in the United States.
The average usage level for air
9.8%
0.001
14.5%
conditioners in my state.
I’m not sure.
56.2%
53.2%
Note: This table reports the results from two qualitative questions we asked at the end of the experiment. The table
replicates the exact wording used for the question and the answers, including underlined text as indicated. We have
excluded a small number of observations (<1%) in which participants refused to answer the question. The correct
answers are highlighted in bold. We calculate all proportions using sampling weights.
41
Table 8: Testing Knowledge About Energy Costs
The national average residential electricity price is 12.4 cents per
kilowatt hour (kWh). How does the average residential electricity
price in your state compare to the national average?
My state’s electricity prices are higher than
14.3%
the national average.
My state’s electricity prices are lower than
16.6%
the national average.
I’m not sure.
69.2%
Percentage Correct
20.2%
How do you think average air conditioning usage in your state
compares to the average usage nationally?
Average usage in my state is probably
30.6%
higher than the national average.
Average usage in my state is probably lower
28.1%
than the national average.
I’m not sure.
41.3%
Percentage Correct
40.4%
Note: This table reports the results from two questions we asked at the end of the
experiment. The table replicates the exact wording used for the question and the
answers, including underlined text as indicated. We have excluded a small
number of observations (<1%) in which participants refused to answer the
question. The percentage correct is the fraction of participants who are able to
answer the question (i.e. they don’t respond “I’m not sure”) and are correct in
how their local conditions compare to the national average. We calculate all
proportions using sampling weights.
42
Table 9. Do Participants Take Local Factors Into Account?
Participants Shown
Participants Shown
Current Labels
State-Specific Labels
(i.e. Control Group)
(i.e. Treatment Group)
(1)
(2)
Electricity Price (cents per
-0.036
0.041*
kWh)
(0.025)
(0.024)
Annual Hours of Air
0.0050
0.0398***
Conditioning Usage
(0.0085)
(0.0085)
(in 100s)
p-value for joint test that
price and usage do not
0.245
0.000
influence EER choice
Number of Observations
3670
3605
Note: This table reports estimated coefficients and standard errors from two separate regressions. For
column (1) the sample is restricted to the 3,670 choices made by participants in the control group and for
column (2) the sample is restricted to the 3,605 choices made by participants in the treatment group. The
dependent variable in both regressions is the energy-efficiency of the selected air conditioner (measured in
EER). In addition to the independent variables listed in the row headings, both regressions include
household income and indicator variables for college graduate, non-white, married, age 65 or over, and
political party affiliation. In both regressions observations are weighted using sampling weights. Standard
errors are clustered by participant.
*** p<0.01, ** p<0.05, * p<0.1
43
Table 10. The Impact of State-Specific Labels on Electricity Consumption
Annual Electricity Consumption
(in kilowatt hours)
Entire Sample
-16.5
(11.1)
10.2
(9.2)
8.5
(16.2)
-64.6**
(27.0)
Low Operating Cost States
Medium Operating Cost States
High Operating Cost States
Note: This table reports coefficient estimates and standard errors corresponding to the
treatment indicator variable from four separate least squares regressions. The dependent
variable in all regressions is the annual electricity consumption in kilowatt hours of the air
conditioner selected by the participant based on annual cooling hours in the state where the
participant lives. All regressions include state fixed effects as well as household income and
indicator variables for college graduate, non-white, married, age 65 or over, and political
party affiliation. For the first row the sample includes all 7,275 choices made by the 2,440
participants in our online experiment. For the regressions reported in the second through
fourth rows, states are divided into three groups (terciles) based on average energy costs
(residential electricity prices multiplied by annual hours of air conditioning use) and then
regressions are run using participants from each subset of states. In all regressions
observations are weighted using sampling weights. Standard errors are clustered by
participant.
*** p<0.01, ** p<0.05, * p<0.1
44
Figure 1: Annual Cooling Hours by State
0 − 500
500 − 1000
1000 − 1500
1500 − 2000
2000 − 2500
2500 − 3000
45
Figure 2: Residential Electricity Prices by State
Cents/kWh
8 − 10
10 − 12
12 − 14
14 − 16
16 − 18
46
Figure 3: The Cost of Operating a Medium-Sized Room Air Conditioner
Dollars Per Year
25 − 100
100 − 175
175 − 250
250 − 325
47
Figure 4: Control and Treatment Labels
Control Group
Treatment Group
48
10.4
Mean Energy−Efficiency of Selected Units (EER)
10.6
10.8
11
11.2
11.4
Figure 5. Do Better Labels Lead to Better Choices?
States with Low Operating Costs
States with Medium Operating Costs
Current Labels
States with High Operating Costs
State−Specific Labels
49
Figure 6. Share of New Air Conditioners Sold in 2009 that are Energy Star
25 − 30
30 − 35
35 − 40
40 − 45
45 − 50
50 − 55
50
Appendix I: Additional Information About the Experiment
The experiment ran from July 16 to August 1, 2014. Participants in the experiment were
not allowed to go back to earlier questions upon moving forward to the next screen in the
experiment. If they did not answer a question, they were prompted once to answer and
then they were allowed to proceed. Less than one percent of responses were blank.
A1 – Complete Survey Instrument
[Screen 1]
Q1. What type of air conditioning equipment do you have in your dwelling?
I have a central air conditioning system. ................................1
I have a room air conditioning unit (or units). ......................2
I don’t have air conditioning in my dwelling. ........................3
[Screen 2]
Imagine that a room has been added to your house that is not cooled by your central air
conditioner.32 You have decided to purchase a room air conditioner for this room.
The next screen will describe three different air conditioners. For each option you will be
shown an Energy Guide label which provides information about operating costs. The
purchase price is also provided for each option. You will be asked to select your preferred
air conditioner.
Assume that all characteristics other than purchase and operating cost of the three
options are identical and that the air conditioner has been properly sized for this room.
Feel free to use a calculator and/or scratch pad to assist you in evaluating the options.
[Screens 3-5]
Q2 – Q4.
32
If the participant does not have central air, the first sentence reads "Imagine a room in your house that is
not currently air conditioned."
51
Note: The screen shot below is a representative set of room air conditioner choices among
which the participant were asked to choose.
[Screen 6]
Q5. What electricity price was used to calculate estimated yearly energy cost in the
Energy Guide labels you were shown?
The average electricity price in the United
States. .................................................................1
The average electricity price in my state. ...........2
I’m not sure. ..........................................................3
[Screen 7]
Q6. The national average residential electricity price is 12.4 cents per kilowatt hour
(kWh). How does the average residential electricity price in your state compare to the
national average?
My state’s electricity prices are higher than the
national average. ..............................................1
My state’s electricity prices are lower than the
national average. ..............................................2
I’m not sure. ..........................................................3
52
[Screen 8]
Q7. Operating costs for an air conditioner depend on the cost of electricity and the
number of hours the air-conditioner is used. What usage level was used to calculate
estimated yearly energy cost in the Energy Guide labels you were shown?
The average usage level for air conditioners in
the United States. .............................................1
The average usage level for air conditioners in
my state. ............................................................2
I’m not sure. ..........................................................3
[Screen 9]
Q8. How do you think average air conditioning usage in your state compares to the
average usage nationally?
Average usage in my state is probably higher
than the national average. ................................1
Average usage in my state is probably lower
than the national average. ...............................2
I’m not sure. ..........................................................3
[Screen 10]
Q9. How do you think air conditioning usage in your home compares to the average
usage in your state?
Usage in my home is probably higher than the
state average. .....................................................1
Usage in my home is probably lower than the
state average. .....................................................2
Usage in my home is probably very close to the
state average. .....................................................3
End of Experiment
A2 –Sampling Design
1. Sample designed to match key benchmarks
53
•
•
•
•
•
•
•
•
•
Gender: Male or Female
Age: 18–29, 30–44, 45–59, and 60+
Race/Ethnicity: Hispanic and non-Hispanic White, Black, Other, and 2+ Races
Education: Less than High School, High School, Some College, Bachelor and
beyond
Census Region: Northeast, Midwest, South, and West
Household Income: $0-$10K, $10K-<$25K, $25K-<$50K, $50K-<$75K, $75K<$100K, and $100K+
Home Ownership: Own or Rent/Other
Metropolitan Area: Yes or No
Home Internet Access: Yes or No
2. Study specific final weights computed to adjust for experiment-specific nonresponse
along the following dimensions
 gender
 race/ethnicity
 education
 census region
 household income
 home internet service
A3 – Appliance Choice Questions
Each participant in the experiment was shown a screen with three room air conditioner
choices and asked to select their most preferred model. The choices ranged from least to
most expensive. More energy-efficient air conditioners were more expensive. After
selecting their preferred model, they were asked to make two additional purchase
decisions in an identical manner to their first selection decision. The design matrix for
the three sets of choices was as follows:
Table A1. Purchase Prices for Air Conditioner Choices
Energy Efficiency Rating (EER)
7.8
11.0
13.1
$200
$420
$550
Decision 1
$200
$505
$600
Decision 2
$200
$335
$440
Decision 3
54
That is, the energy-efficiency rating of the three choices was the same for all participants
and all decisions. We varied the purchase prices across decisions as indicated above. For
all three questions, participants were shown labels that calculated energy costs as a
function of the EER rating (7.8, 11.0, or 13.1) and either average national usage and
electricity prices (control group) or average state usage and electricity prices. Annual
energy costs (EC) are given by the following formula:
∙
(A1)
where BTU is the rated size of the air conditioner (10,000 BTU's for our experiment).
We selected these EER values to reflect typical levels for room air conditioners in the
market for sale in 2014. Then we selected the purchase prices based on simulation
evidence to maximize the precision of our estimates. In particular, we constructed
synthetic nationally-representative data, and then for an assumed distribution of discount
rates simulated choices using draws from an extreme value distribution. We included
both a treatment group and a control group and assumed both would take the
information in the labels at face value. Then with the generated “data,” we estimated the
discount rate and examined the distribution of choices across states as in Figure 5 from
the paper. The purchase prices above were those that minimized the standard error of the
estimated discount rate and provided a good mix of choices across states for both the
treatment and control groups.
55